So I was debugging a large binary memory usage in my application. Using :recon I found that two geolix processes were holding onto large binary references. I'm using the MMDB2 adapter. Before these changes, my BEAM instance was using ~800MB of memory.
After modifying the loader processes to hibernate and trigger a garbage collection, the memory usage went down to ~270MB.

This is the snippet using recon to find processes holding onto binaries:

Process.list() |> Enum.map(fn pid -> {pid, :recon.info(pid, :binary_memory) |> elem(1) |> then(fn x -> x / 1024 / 1024 end)} end) |> Enum.sort_by(fn {_pid, mem} -> mem end, :desc) |> Enum.take(4)

Resulting in:

[
  {#PID<0.766.0>, 531.5929861068726},
  {#PID<0.774.0>, 234.94092750549316},
  {#PID<0.590.0>, 4.768865585327148},
  {#PID<0.1678.0>, 4.103562355041504}
]

The geolix loader (in this case 0.766.0) was holding onto ~530MB of binaries while the mmdb2 loader was holding onto ~234MB.

There is a similar PR for the mmdb2 adapter. If you know the proper places where the mmdb2 adapter creates the binary references, another fix would be to call :binary.copy there instead. I don't think that hibernating here has any bad consequences though.